Independent AI product · Piaopiao Travel
A Chinese-language outbound-travel assistant that plans trips through an AI chat and maintains entry-requirement data for 199 countries and regions. The product rule is simple: the model writes the plan, but every entry requirement comes from checked data, not from the model.
This page collects the decision records behind the product rather than its feature list: how a stalled funnel was diagnosed from server-side data, how a 199-country dataset is governed, how an AI feature is measured before it ships, and how a model decision was made on cost, latency, and quality.
Data-governance and evaluation figures as of 2026-10-28. Basis and sources for each are in the evidence appendix.
199
Countries and regions covered
91
With verified official sources (261 units, 388 sources)
366 + 55
Official URLs and policy-page fingerprints checked weekly
122
Change-log entries, reproduced byte for byte in CI
200 × 4
Held-out samples × models on one extraction benchmark
25.2 s
Itinerary generation, down from 94 s
As of 2026-10-28 · Basis and sources
Each one starts from a business question and ends with what the evidence supports, including what it does not.
Sign-ups stopped for 55 days and there was no page analytics. What can the database alone explain?
Sixteen read-only aggregate queries, stated metric definitions, and a clear line between the one confirmed cause, the hypotheses it ruled out, and the ones that still need a test.
How does every fact trace back to an official page and the date it was checked?
Sourcing, weekly change detection, a traceable change log, alert-to-action routing, and coverage reported as it actually is.
Is it good now, did this change make it better, and should it ship?
Layered metrics instead of one overall score, fixed baselines, release gates, and the cases where the measurement itself misleads.
Keep the vendor model, switch vendors, or train our own on distilled data?
Four models on the same 200 held-out samples, with latency and cost per call, judged against rules written in advance. The fine-tuned model was evaluated, not deployed.
The data-governance and evaluation figures cited in the case studies, each with its as-of date, basis, and source.